17582_02_ch02_p015-054.qxd 24 11/25/08 3:27 PM Page 24 Chapter 2 Using Surveys and Experimental Studies to Gather Data 2.3 Sampling Designs for Surveys A crucial element in any survey is the manner in which the sample is selected from the population. If the individuals included in the survey are selected based on convenience alone, there may be biases in the sample survey, which would prevent the survey from accurately reflecting the population as a whole. For example, a marketing graduate student developed a new approach to advertising and, to evaluate this new approach, selected the students in a large undergraduate business course to assess whether the new approach is an improvement over standard advertisements. Would the opinions of this class of students be representative of the general population of people to which the new approach to advertising would be applied? The income levels, ethnicity, education levels, and many other socioeconomic characteristics of the students may differ greatly from the population of interest. Furthermore, the students may be coerced into participating in the study by their instructor and hence may not give the most candid answers to questions on a survey. Thus, the manner in which a sample is selected is of utmost importance to the credibility and applicability of the study’s results. In order to precisely describe the components that are necessary for a sample to be effective, the following definitions are required. target population sample sampled population observation unit sampling unit Target population: The complete collection of objects whose description is the major goal of the study. Designating the target population is a crucial but often difficult part of the first step in an observational or experimental study. For example, in a survey to decide if a new storm-water drainage tax should be implemented, should the target population be all persons over the age of 18 in the county, all registered voters, or all persons paying property taxes? The selection of the target population may have a profound effect on the results of the study. Sample: A subset of the target population. Sampled population: The complete collection of objects that have the potential of being selected in the sample; the population from which the sample is actually selected. In many studies, the sampled population and the target population are very different. This may lead to very erroneous conclusions based on the information collected in the sample. For example, in a telephone survey of people who are on the property tax list (the target population), a subset of this population may not answer their telephone if the caller is unknown, as viewed through caller ID. Thus, the sampled population may be quite different from the target population with respect to some important characteristics such as income and opinion on certain issues. Observation unit: The object upon which data are collected. In studies involving human populations, the observation unit is a specific individual in the sampled population. In ecological studies, the observation unit may be a sample of water from a stream or an individual plant on a plot of land. Sampling unit: The object that is actually sampled. We may want to sample the person who pays the property tax but may only have a list of telephone numbers. Thus, the households in the sampled population serve as the sampled units, and the observation units are the individuals residing in the sampled household. In an entomology study, we may sample 1-acre plots of land and then count the number of insects on individual plants 17582_02_ch02_p015-054.qxd 11/25/08 3:27 PM Page 25 2.3 Sampling Designs for Surveys sampling frame simple random sampling stratified random sample ratio estimation cluster sampling 25 residing on the sampled plot. The sampled unit is the plot of land, the observation unit would be the individual plants. Sampling frame: The list of sampling units. For a mailed survey, it may be a list of addresses of households in a city. For an ecological study, it may be a map of areas downstream from power plants. In a perfect survey, the target population would be the same as the sampled population. This type of survey rarely happens. There are always difficulties in obtaining a sampling frame or being able to identify all elements within the target population. A particular aspect of this problem is nonresponse. Even if the researcher was able to obtain a list of all individuals in the target population, there may be a distinct subset of the target population which refuses to fill out the survey or allow themselves to be observed. Thus, the sampled population becomes a subset of the target population. An attempt at characterizing the nonresponders is very crucial in attempting to use a sample to describe a population. The group of nonresponders may have certain demographics or a particular political leaning that if not identified could greatly distort the results of the survey. An excellent discussion of this topic can be found in the textbook, Sampling: Design and Analysis by Sharon L. Lohr (1999), Pacific Grove, CA: Duxbury Press. The basic design (simple random sampling) consists of selecting a group of n units in such a way that each sample of size n has the same chance of being selected. Thus, we can obtain a random sample of eligible voters in a bond-issue poll by drawing names from the list of registered voters in such a way that each sample of size n has the same probability of selection. The details of simple random sampling are discussed in Section 4.11. At this point, we merely state that a simple random sample will contain as much information on community preference as any other sample survey design, provided all voters in the community have similar socioeconomic backgrounds. Suppose, however, that the community consists of people in two distinct income brackets, high and low. Voters in the high-income bracket may have opinions on the bond issue that are quite different from the opinions of low-income bracket voters. Therefore, to obtain accurate information about the population, we want to sample voters from each bracket. We can divide the population elements into two groups, or strata, according to income and select a simple random sample from each group. The resulting sample is called a stratified random sample. (See Chapter 5 of Scheaffer et al., 2006.) Note that stratification is accomplished by using knowledge of an auxiliary variable, namely, personal income. By stratifying on high and low values of income, we increase the accuracy of our estimator. Ratio estimation is a second method for using the information contained in an auxiliary variable. Ratio estimators not only use measurements on the response of interest but they also incorporate measurements on an auxiliary variable. Ratio estimation can also be used with stratified random sampling. Although individual preferences are desired in the survey, a more economical procedure, especially in urban areas, may be to sample specific families, apartment buildings, or city blocks rather than individual voters. Individual preferences can then be obtained from each eligible voter within the unit sampled. This technique is called cluster sampling. Although we divide the population into groups for both cluster sampling and stratified random sampling, the techniques differ. In stratified random sampling, we take a simple random sample within each group, whereas in cluster sampling, we take a simple random sample of groups and then sample all items within the selected groups (clusters). (See Chapters 8 and 9 of Scheaffer et al., 2006, for details.) 17582_02_ch02_p015-054.qxd 26 11/25/08 3:27 PM Page 26 Chapter 2 Using Surveys and Experimental Studies to Gather Data systematic sample Sometimes, the names of persons in the population of interest are available in a list, such as a registration list, or on file cards stored in a drawer. For this situation, an economical technique is to draw the sample by selecting one name near the beginning of the list and then selecting every tenth or fifteenth name thereafter. If the sampling is conducted in this manner, we obtain a systematic sample. As you might expect, systematic sampling offers a convenient means of obtaining sample information; unfortunately, we do not necessarily obtain the most information for a specified amount of money. (Details are given in Chapter 7 of Scheaffer et al., 2006.) The following example will illustrate how the goal of the study or the information available about the elements of the population determine which type of sampling design to use in a particular study. EXAMPLE 2.3 Identify the type of sampling design in each of the following situations. a. The selection of 200 people to serve as potential jurors in a medical malpractice trial is conducted by assigning a number to each of 140,000 registered voters in the county. A computer software program is used to randomly select 200 numbers from the numbers 1 to 140,000. The people having these 200 numbers are sent a postcard notifying them of their selection for jury duty. b. Suppose you are selecting microchips from a production line for inspection for bent probes. As the chips proceed past the inspection point, every 100th chip is selected for inspection. c. The Internal Revenue Service wants to estimate the amount of personal deductions taxpayers made based on the type of deduction: home office, state income tax, property taxes, property losses, and charitable contributions. The amount claimed in each of these categories varies greatly depending on the adjusted gross income of the taxpayer. Therefore, a simple random sample would not be an efficient design. The IRS decides to divide taxpayers into five groups based on their adjusted gross incomes and then takes a simple random sample of taxpayers from each of the five groups. d. The USDA inspects produce for E. coli contamination. As trucks carrying produce cross the border, the truck is stopped for inspection. A random sample of five containers is selected for inspection from the hundreds of containers on the truck. Every apple in each of the five containers is then inspected for E. coli. Solution a. A simple random sample is selected using the list of registered voters as the sampling frame. b. This is an example of systematic random sampling. This type of inspection should provide a representative sample of chips because there is no reason to presume that there exists any cyclic variation in the production of the chips. It would be very difficult in this situation to perform simple random sampling because no sampling frame exists. c. This is an example of stratified random sampling with the five levels of personal deductions serving as the strata. Overall the personal deductions of taxpayers increase with income. This results in the stratified random 17582_02_ch02_p015-054.qxd 11/25/08 3:27 PM Page 27 2.3 Sampling Designs for Surveys 27 sample having a much smaller total sample size than would be required in a simple random sample to achieve the same level of precision in its estimators. d. This is a cluster sampling design with the clusters being the containers and the individual apples being the measurement unit. The important point to understand is that there are different kinds of surveys that can be used to collect sample data. For the surveys discussed in this text, we will deal with simple random sampling and methods for summarizing and analyzing data collected in such a manner. More complicated surveys lead to even more complicated problems at the summarization and analysis stages of statistics. The American Statistical Association (http://www.amstat.org) publishes a series of brochures on surveys: What Is a Survey? How to Plan a Survey, How to Collect Survey Data, Judging the Quality of a Survey, How to Conduct Pretesting, What Are Focus Groups? and More about Mail Surveys. These describe many of the elements crucial to obtaining a valid and useful survey. They list many of the potential sources of errors commonly found in surveys with guidelines on how to avoid these pitfalls. A discussion of some of the issues raised in these brochures follows. Problems Associated with Surveys survey nonresponse Even when the sample is selected properly, there may be uncertainty about whether the survey represents the population from which the sample was selected. Two of the major sources of uncertainty are nonresponse, which occurs when a portion of the individuals sampled cannot or will not participate in the survey, and measurement problems, which occur when the respondent’s answers to questions do not provide the type of data that the survey was designed to obtain. Survey nonresponse may result in a biased survey because the sample is not representative of the population. It is stated in Judging the Quality of a Survey that in surveys of the general population women are more likely to participate than men; that is, the nonresponse rate for males is higher than for females. Thus, a political poll may be biased if the percentage of women in the population in favor of a particular issue is larger than the percentage of men in the population supporting the issue. The poll would overestimate the percentage of the population in favor of the issue because the sample had a larger percentage of women than their percentage in the population. In all surveys, a careful examination of the nonresponse group must be conducted to determine whether a particular segment of the population may be either under- or overrepresented in the sample. Some of the remedies for nonresponse are 1. Offering an inducement for participating in the survey 2. Sending reminders or making follow-up telephone calls to the individuals who did not respond to the first contact 3. Using statistical techniques to adjust the survey findings to account for the sample profile differing from the population profile measurement problems Measurement problems are the result of the respondents not providing the information that the survey seeks. These problems often are due to the specific wording of questions in a survey, the manner in which the respondent answers the survey